This article was downloaded by: [103.141.127.

88] On: 18 May 2024, At: 10:33

Publisher: Institute for Operations Research and the Management Sciences (INFORMS)
INFORMS is located in Maryland, USA

Manufacturing & Service Operations Management

Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org

Logistics Performance, Ratings, and Its Impact on

Customer Purchasing Behavior and Sales in E-Commerce
Platforms
Vinayak Deshpande, Pradeep K. Pendem

To cite this article:

Vinayak Deshpande, Pradeep K. Pendem (2023) Logistics Performance, Ratings, and Its Impact on Customer Purchasing
Behavior and Sales in E-Commerce Platforms. Manufacturing & Service Operations Management 25(3):827-845. https://
doi.org/10.1287/msom.2021.1045

Full terms and conditions of use: https://pubsonline.informs.org/Publications/Librarians-Portal/PubsOnLine-Terms-and-

Conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.

Copyright © 2022, INFORMS

Please scroll down for article—it is on subsequent pages

With 12,500 members from nearly 90 countries, INFORMS is the largest international association of operations research (O.R.)
and analytics professionals and students. INFORMS provides unique networking and learning opportunities for individual
professionals, and organizations of all types and sizes, to better understand and use O.R. and analytics tools and methods to
transform strategic visions and achieve better outcomes.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
 MANUFACTURING & SERVICE OPERATIONS MANAGEMENT
Vol. 25, No. 3, May–June 2023, pp. 827–845
https://pubsonline.informs.org/journal/msom ISSN 1523-4614 (print), ISSN 1526-5498 (online)

Finalist---2018 MSOM Data Driven Research Challenge

Logistics Performance, Ratings, and Its Impact on Customer

Purchasing Behavior and Sales in E-Commerce Platforms
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

Vinayak Deshpande,a Pradeep K. Pendemb

a
Kenan-Flagler Business School, University of North Carolina, Chapel Hill, North Carolina 27599; b Charles H. Lundquist College of Business,
University of Oregon, Eugene, Oregon 97403
Contact: vinayak_deshpande@kenan-flagler.unc.edu, https://orcid.org/0000-0002-6681-028X (VD); pradeepp@uoregon.edu,
https://orcid.org/0000-0001-7900-9427 (PP)

Received: February 17, 2019 Abstract. Problem definition: We examine the impact of logistics performance metrics such as
Revised: September 21, 2020; May 17, 2021 delivery time and customer’s requested delivery speed on logistics service ratings and third-
Accepted: July 9, 2021 party sellers’ sales on an e-commerce platform. Academic/practical relevance: Although
Published Online in Articles in Advance: e-commerce retailers like Amazon have recently invested heavily in their logistics networks to
January 11, 2022 provide faster delivery to customers, there is scant academic literature that tests and quantifies
https://doi.org/10.1287/msom.2021.1045 the premise that convenient and fast delivery will drive sales. In this paper, we provide empiri-
cal evidence on whether this relationship holds in practice by analyzing a mechanism that con-
Copyright: © 2022 INFORMS nects delivery performance to sales through logistics ratings. Prior academic work on online rat-
ings in e-commerce platforms has mostly analyzed customers’ response to product functional
performance and biases that exist within. Our study contributes to this stream of literature by
examining customer experience from a service quality perspective by analyzing logistics service
performance, logistics ratings, and its impact on customer purchase probability and sales.
Methodology: Using an extensive data set of more than 15 million customer orders on the Tmall
platform and Cainiao network (logistics arm of Alibaba), we use the Heckman ordered regres-
sion model to explain the variation in customers’ rating of logistics performance and the likeli-
hood of customers posting a logistics rating. Next, we develop a generic customer choice model
that links the customer’s likelihood of making a purchase to the logistics ratings provided by
prior customers. We implement a two-step estimation of the choice model to quantify the im-
pact of logistics ratings on customer purchase probability and third-party seller sales. Results:
We surprisingly find that even customers with no promise on delivery speed are likely to post
lower logistics ratings for delivery times longer than two days. Although these customers are
not promised an explicit delivery deadline, they seem to have a mental threshold of two days
and expect deliveries to be made within that time. Similarly, we find that priority customers
(those with two-day and one-day promise speed) provide lower logistics ratings for delivery
times longer than their anticipated delivery date. We estimate that reducing the delivery time
of all three-day delivered orders on this platform (which makeup ≈ 35% of the total orders) to
two days would improve the average daily third-party seller sales by 13.3% on this platform.
The impact of delivery time performance on sales is more significant for sellers with a higher
percentage of three-day delivered orders and a higher spend per order. Managerial implica-
tions: Our study emphasizes that delivery performance and logistics ratings, which measure
service quality, are essential drivers of the customer purchase decision on e-commerce plat-
forms. Furthermore, by quantifying the impact of delivery time performance on sales, our study
also provides a framework for online retailers to assess if the increase in sales because of im-
proved logistics performance can offset the increase in additional infrastructure costs required
for faster deliveries. Our study’s insights are relevant to third-party sellers and e-commerce
platform managers who aim to improve long-term online customer traffic and sales.

History: This paper has been accepted as part of the 2018 MSOM Data Driven Research Challenge.
Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2021.1045.

Keywords: e-commerce • integrated warehousing and logistics • logistics ratings • big data analytics

1. Introduction retailers provide several advantages for a customer

E-commerce is one of the largest growing sectors in over traveling to a brick-and-mortar retail store, such
the digital economy, with sales of $766.77 billion in as a more comprehensive selection of products and
the United States in 2019 and compound growth of the ability to find desired products quickly. With the
20.8% during 2014-2019 (IBISWorld 2019). Online explosion of digitization and the added convenience

827
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
828 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

of access to mobile devices, the industry expects to faster shipping has given an additional boost to sales.
reach sales of $1,916 billion in 2024 with compounded In this paper, we focus on quantifying the value of
growth of 20.1% from 2019 to 2024 (IBISWorld 2019). fast shipping by examining the impact of delivery
With the belief that fast and convenient delivery time reduction on the online platform’s sales. In par-
will attract more new customers, online retailers ticular, we provide empirical evidence on whether
started to offer premium membership subscription, this relationship holds in practice by analyzing a
which typically promises fast shipping. For example, mechanism that connects delivery performance to
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

Amazon launched a premium service in 2005, Prime, sales through logistics ratings.
with a promise of two-day shipping as the main From a customer’s perspective, products on most
draw. Fast forward to 2020; the company now has online platforms are sold by two kinds of sellers: (i)
more than 150 million customers around the world The platform itself as the seller. For example, Amazon
who subscribe to Prime membership service (Del Rey sells products on its platform, (ii) Independent third-
2020). Amazon has continued innovating and inves- party sellers who sell on the platform. We channel our
ting significantly in fast shipping advancements to im- effort to particularly examine the third-party sellers as
prove customer service experience. According to Jeff they have proliferated in recent years. The proportion
Bezos, the number of items delivered to U.S. custom- and share of revenue of third-party sellers on Amazon
ers with Prime’s free one-day and same-day delivery have grown from 25% to 52% during 2007–2018
more than quadrupled during 2018–2019 (Del Rey (Richter 2019, Clement 2020). This leads to our main
2020). A similar trend toward fast shipping services research question: What is the impact of fast shipping
has spread across some of the U.S.’s top retailers, such or reducing delivery time on sales for a third-party
as Walmart, Target, and Costco (Montasell 2020). For seller on the e-commerce platform? (Q1). We identify
example, in response to competition from Amazon, the mechanism behind the relationship between deliv-
Walmart launched a two-day shipping service in June ery performance and sales in two phases. In the first
2016, and three years later, it started offering next-day phase, we establish the link between delivery perfor-
shipping in May 2019 (Crook 2016, Perez 2019). Simi- mance and online logistics ratings, and in the second
larly, Target launched free two-day and same-day phase, we link logistics ratings to online sales, as de-
deliveries in October 2018 and June 2019, respectively scribed later.
(Perez 2018, Liptak 2019). Overall, market trends In the current digital age, sellers can now offer a
show that big online retailers invest heavily to im- wide variety of products and different versions (e.g.,
prove the customer service experience by reducing de- model, style, color) of a single product. The rich infor-
livery time. mation available about products means that custom-
Although faster shipping service is convenient and ers can choose from a wide range of options. To ease
provides superior service experience to customers, it the search cost, reduce product uncertainty (Chen and
has significant cost implications for online retailers. The Xie 2008), and help customers in the purchase decision
additional capacity investments required to extend the process, most platforms provide ratings of products
reach of faster delivery service range from expanding and sellers in addition to information such as price,
fulfillment centers, distribution centers, air-cargo opera- discounts, and available inventory. The ratings are
tions to the rental fleet. For example, Amazon spent typically displayed in the form of a distribution rang-
$21.7 billion on shipping costs in 2017, nearly twice ing from 1 to 5, with 1 being the lowest quality and 5
what it spent in 2015 (Semuels 2018). Amazon’s world- being the highest quality. Sellers rely heavily on rat-
wide shipping costs jumped 46% to $9.6 billion in the ings to maintain market share and survive against
third quarter of 2019 from the previous year related to fierce price competition. Platforms typically display
its one-day shipping program for Prime subscribers ratings on two dimensions on quality: product and
(Semuels 2018). The company profits slipped by 26% in logistics service. Product quality signifies the response
October 2019 because of high shipping costs on one-day to customer’s experiences on the product perfor-
deliveries for Prime customers (Mattioli 2019). mance, as stated on their web page. On the other
With increasing e-commerce demand and advance- hand, the logistics service quality typically signifies
ments in fast shipping, online retailers are likely to in- the response to the customer experience of timely or-
cur higher shipping costs in the future. As a result, it der delivery. A customer with an intent to purchase
is of primary interest to online retailers such as Ama- visits the platform obtains price, ratings, and other
zon to understand if the increasing costs incurred in sources of information on the product page. The cus-
fast shipping provide a significant return in additional tomer then evaluates price and ratings across multiple
sales. Although there is anecdotal evidence that online sellers and finally decides on product purchase. At the
sales growth has accelerated with more investments end of the purchase process, sellers generally provide
in one-day shipping (Herrera and Qian 2019), there is different shipping speed options: (i) fast delivery or
scant academic literature and no formal evidence that (ii) regular delivery. At this point, the customer
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 829

decides on the choice of shipping speed and places examine the impact of improved logistics ratings on
the order. The selection of shipping speed by the cus- customer purchase probability and provide estimation
tomer sets an expectation of logistics service quality results. In Section 6, we use the results derived from
for the seller. The customer later receives the product Sections 4 and 5 to quantify the impact of reducing de-
and evaluates both the actual product and logistics livery time on seller sales. In Section 7, we examine
service quality. If the product functionality conforms the robustness of the causal link between delivery per-
to the web page’s specification, the customer likely formance and customer-provided logistics rating by
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

provides a high product quality rating (Chintagunta addressing one form of independent variable endoge-
et al. 2010, Lin et al. 2011). On the other hand, when neity. Last, in Section 8, we conclude our work with
the seller delivers the product either on-time or earli- managerial insights and highlight the limitations of
er, the customer is likely to be elated with the logistics our study.
service experience. However, it is unclear if their re-
sponse after a satisfying service experience manifests 2. Literature
in a higher logistics rating. This leads to the first phase Our work contributes to the following four areas of re-
of our main research question: What is the impact of search: (i) relationship between logistics performance
delivery time on the logistics rating provided by the and financial outcomes, (ii) drivers of ratings on online
customer? (Q1a). platforms, (iii) effect of disconfirmation on customer
Higher ratings about the seller and their products satisfaction, and (iv) customer choice models under
are likely to increase the likelihood of an incoming
censored choices and limited (or no) market size infor-
customer’s purchase. Therefore, it is no surprise that
mation. We discuss the prior literature in each of these
e-commerce platforms make enormous efforts by send-
areas and identify the contribution of our paper.
ing emails or text messages requesting feedback about
their products and service from previous customers. A
Relationship Between Logistics Performance and
cumulative volume of higher logistics ratings from
Financial Outcomes. Recent literature in operations
customers over time increases the effective rating visible
management has looked at empirical evidence on the
to an incoming customer. A higher overall logistics rat-
relationship between logistics performance and finan-
ing is then likely to increase the likelihood of a purchase,
cial outcomes. Allon et al. (2011) estimate that a seven-
which affects the seller’s long-term traffic and sales. Prior
second reduction in customer’s wait time at a fast-food
work on online product quality ratings has shown evi-
drive-thru chain can result in an average 1%–3% in-
dence of their impact on customer purchase probability
or sales (Chevalier and Mayzlin 2006, Chintagunta et al. crease in market share. Cui et al. (2019) finds that re-
2010, Lin et al. 2011). However, the answer to whether a moving a high-quality logistics carrier option for a
higher logistics rating impacts customer purchase deci- large online retailer leads to a decrease in sales by
sions is unclear. This forms the basis for the second 16.42%, whereas its resumption increases by 18.83%.
phase of our main research question: What is the impact Fisher et al. (2019) extrapolate how faster delivery af-
of improved logistics ratings on customer purchase prob- fects sales in the online channel of a U.S. apparel retail-
ability and sales for the seller? (Q1b).1 er by leveraging a quasi-experiment involving the
In summary, we analyze the mechanism by examin- opening of a new distribution center, which resulted
ing the following two research questions individually. in unannounced faster deliveries to the western US
First, what is the impact of reducing delivery time on states through its online channel. Unlike Fisher et al.
the logistics rating provided by a customer? (Q1a). Sec- (2019), who used aggregated average estimates of de-
ond, what is the impact of improved logistics ratings livery time, we use actual observed delivery times to
on customer purchase probability and sales for the individual customers to identify their response to de-
seller? (Q1b). We then combine our results from Q1a livery time performance. The mechanism of improved
and Q1b to answer our main research question: What financial outcomes from superior logistics perfor-
is the impact of reducing delivery time on sales for a mance can be a result of a customer’s response to (i)
third-party seller on the e-commerce platform? (Q1). their own recent or past service experience or (ii) to ac-
The rest of the paper is organized as follows. In Sec- cumulated feedback from previous customers’ experi-
tion 2, we provide details on prior work on logistics ence on the common platform (word-of-mouth). All
performance and online ratings and highlight our con- the studies mentioned attribute the improvement to
tribution to the literature. In Section 3, we briefly de- the first mechanism (customer’s own prior experience)
scribe the study setting, Cainiao’s logistics operations, but do not empirically validate this mechanism. In a
and summarize the data. In Section 4, we provide the recent study, Mao et al. (2019) find that a 10-minute
measures, model, and results on the impact of reduc- earlier delivery from the targeted (or expected) deliv-
ing delivery time on the logistics rating of a seller. ery time increases each customer’s future demand by
In Section 5, we build a generic choice model to 1.03 orders per month on an on-demand meal delivery
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
830 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

service platform. Like the above studies, this work at- service quality affects customer satisfaction. Disconfir-
tributes the increase in sales from superior delivery mation is the discrepancy between customer expecta-
performance to the first mechanism and moreover, tion and the actual performance of product or service
provides an empirical validation of this mechanism. quality. Each customer builds a certain level of expec-
Unlike the prior studies, our research focuses on the tation for the product or quality before initiating the
second mechanism (word-of-mouth effect through lo- purchase. After receiving the product, the customer
gistics ratings) by providing an empirical validation of consumes it and experiences the performance of its
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

this mechanism. In summary, our study contributes to quality. If the performance is worse than expected, the
this stream of literature on quantifying the impact of customer experiences more discomfort, which leads to
faster delivery on financial outcomes by analyzing ac- more dissatisfaction, thereby resulting in lower ratings.
tual delivery times on individual orders to identify the In online platforms, the customer’s actual expectation of
mechanism of customer response to accumulated feed- quality is not explicitly specified by the customer and is
back on delivery performance by other customers. unknown to the seller. Similarly, the performance after
consuming the product is only observed by the custom-
Drivers of Ratings on Online Platforms. The work on er and is unknown to the seller. The response to perfor-
online ratings in the marketing or information systems mance compared with their expectation manifests in
literature has mostly focused on physical products (e.g., the form of a customer’s rating. As a result, research
books, DVDs, and videos). The ratings for these prod- that examined the impact of disconfirmation on cus-
ucts on platforms signify product functional quality. tomer satisfaction has considered the distribution of rat-
Most of the prior work on ratings has primarily focused ings before the purchase, and the rating posted by a
on different biases influencing the ratings/score. These customer as a proxy for expectation and the disconfir-
biases can originate in various forms ranging from the mation (Anderson and Sullivan 1993, Bhattacherjee
level of an individual customer to the firm. Some of the 2001, McKinney et al. 2002, Susarla et al. 2006, Ho et al.
biases that arise at a personal level include self- 2017b). A different stream of literature on online ratings
presentational behavior (Schlosser 2005), self-selection also notes that the distribution is often driven by differ-
(Li and Hitt 2008, Hu et al. 2009), previous opinions ent biases, such as purchasing bias and under-reporting
(Wu and Huberman 2008), and high-quality sellers (Li bias (Hu et al. 2009). As a result, the measure of discom-
and Xiao 2013). Social biases comprise identity (Wang fort based on the prior distribution of ratings and its
2010), comparisons (Chen et al. 2010), high valance, and impact on customer satisfaction is likely biased.
high volume rating environments (Moe and Schweidel Like Bray (2020), our study contributes to this stream
2012), friends and strangers (Lee et al. 2015), and cultur- of literature by proposing an accurate measure of dis-
al influences (Koh et al. 2010). At the organizational lev- comfort and its impact on customer satisfaction. In our
el, Dellarocas (2006) showed that ratings might not be study, a customer explicitly states their expectations
wholly trusted as they could be strategically manipulat- (choice of shipping speed) and performance (delivery
ed by firms (potentially inflated) to remain competitive time) is observed both by the customer and seller.
or maintain market share. Hence, the disconfirmation derived from the selection
Our work differs from previous research as we study of shipping speed and delivery time generates an accu-
logistics service quality rather than product quality. rate measure of discomfort. Furthermore, our study
Specifically, the ratings that we consider are the cus- also contributes to understanding the effects of discom-
tomer’s response to an experience of logistics service- fort on a customer’s likelihood of posting a rating by
fulfillment from the point of purchase to delivery rather different customer types. Ho et al. (2017b) finds that an
than product performance quality. The closest work on individual is more likely to post a review when the
logistics ratings in operations management is by Bray magnitude of disconfirmation (s)he encounters is larger.
(2020), where the author shows that scores are higher Our study also finds a similar result for customers with
when the track-package activities cluster toward the no promise on delivery speed and, hence, extends our
end of the shipping horizon than at the start. Our con- understanding of the effect of discomfort for customers.
tribution to this stream of literature is primarily on ex-
amining the impact of observable (both to the seller and Customer Choice Models Under Censored Choices
customer) operational factors on logistics ratings, which and Limited (or No) Market Size Information.
has not been studied previously. Customer choice models describe the decision mak-
er’s choices among alternatives derived under the as-
Effect of Disconfirmation on Customer Satisfaction. sumption of utility-maximizing behavior (Train 2009).
The expectation-disconfirmation theory (Oliver 1977, The alternatives (referred to as the choice set) repre-
1980) in service marketing is a prominent theory ap- sent competing products or sellers over which choices
plied to understand drivers of customer satisfaction. are made. The modeling framework allows the choice
The theory states that disconfirmation of product or decision to be related to explanatory variables. The
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 831

simplicity of these models allows for the derivation of stated in the paper. A brief description of the study
choice probability expressions without reference to setting is as follows.
how the choice is made. The model parameters are The Cainiao Network is a consortium of warehouse
then estimated by equating the choice probabilities to and carrier companies founded by Alibaba. The net-
observed market shares based on sales information. work operates as integrated warehousing and logistics
The simpleness of the framework and the expressions platform linking storage, distribution centers, and carri-
comes with critical information requirements without er providers. Tmall is a business-to-consumer (B2C)
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

which estimation can result in biased and inconsistent Chinese language online retail website operated by
parameters. One such information is the availability Alibaba since April 2008. As of February 2018, the site
of an exhaustive list of competitors, sales, price, and had more than 500 million monthly active users. Before
market size information. Often because of data limita- forming the network in 2013, sellers on Tmall managed
tions, only a subset of such information is available. the inventory and fulfilled orders by themselves or by
For example, when an exhaustive list of competitors is using various third-party carriers. The Cainiao network
unavailable or partly available, the general approach opened its service to all the sellers on their platform
is to put those competitors into an outside option who now have the option to have their inventory man-
(Vulcano et al. 2012, Newman et al. 2014) or use com- aged and customer orders fulfilled conveniently.
petitor stockouts and price variations (Fisher et al. The life cycle of a customer order on the Tmall plat-
2018). Similarly, when market size (number of poten- form from its inception to delivery is as follows. A
tial customers) information is unavailable, the ap- customer visits the platform on a personal computer or a
proach is to develop expressions free of market size mobile app. Next, (s)he browses the products from dif-
by integrating across all possible values (Vulcano et al. ferent sellers and places the order for a seller on the plat-
2012, Newman et al. 2014) or assuming a market size form (“order arrival” event in our data). Sellers on the
based on demographic information (Berry et al. 1995). platform are classified into two types based on whether
In our context, we face the same challenges where we they depend on the Cainiao network for management of
have data on products, their attributes, and sales infor- product inventory and fulfillment of order delivery. The
mation only from sellers operating on a single platform. first type of sellers is independent of Cainiao and man-
It is likely possible that an incoming customer can access age the inventory of products in their warehouse. These
sellers are responsible for shipping the product from
similar products from other channels such as competing
their warehouse to the carrier location. The second type
online platforms (e.g., JD.com) and brick-and-mortar
of sellers is dependent on Cainiao for managing the
stores. In the standard Multinomial Logit (MNL) model
product inventory. Cainiao is responsible for shipping
(Train 2009), the customer compares the utility received
these products to the carrier location. The package is
from different alternatives and decides to choose one of
next shipped (“consignment” event in our data) to the
them or do not purchase at all (typically referred to as
carrier location. The carrier then sends the package via
the “outside option,” and its utility is assumed to be
multiple logistics transfers (e.g., air, rail, and road) to the
zero). Because the sellers on the platform that we study
delivery station near the customer location. The package
do not form an exhaustive list of competitors, customers
is then finally shipped from the delivery station to the
can receive nonzero utility of purchase from other chan-
customer’s location by the carrier. The customer ac-
nels, thus violating the assumption of zero utility for the knowledges receipt of the package by providing a signa-
outside option. We use the framework in Newman et al. ture (“signed” event in our data) to the delivery person.
(2014), which proposes an efficient and consistent esti- Later, the customer may decide to post a rating for their
mation of choice-based models when one alternative is experience on the platform, including their experience
completely censored. Other channels of purchase, such on order quality, online purchase experience, and the lo-
as competing online platforms and brick-and-mortar gistics service quality.
stores, are entirely censored in our context. We build We use information on customer orders, a detailed
on this choice model and extend it to incorporate infor- timeline of delivery events for each of these individual
mation on actual market size through the number of orders, and information about unique customer visits
unique customer visits to the platform. Furthermore, to the platform from April 2017 for our analysis. The
we allow utility specification of the “no-purchase” al- customer order data possess granular information in-
ternative to vary by time, which has not been consid- cluding the date and time-stamp of order arrival, cus-
ered in the prior literature. tomer identifier, item identifier, quantity ordered, to-
tal payment in Chinese yuan, customer’s requested
3. Study Setting and Data delivery speed (can take three unique values:
We use data from the Cainiao network and the Tmall no-promise speed, two-day promise speed, and one-
platform provided by the 2018 MSOM data-driven day promise speed), Cainiao indicator (1: if the order
research challenge to study the research questions used Cainiao warehousing and logistics service, 0:
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
832 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

otherwise), seller identifier, carrier identifier, and 4.1. Measures

logistics rating (customer’s response to the logistics Using each customer order as the unit of analysis, we
service, uses a Likert scale from 1 to 5 with 1 being aim to understand the impact of different operational
low quality and 5 being high quality). variables such as delivery time, total order payment,
The logistics data describes the event timeline from promise speed of delivery, and additional variables
warehouse shipment to final delivery for each custom- on the logistics rating. The detailed definitions for
er order. The time duration from the order arrival each of these measures, which go into the econometric
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

event to the signed event measures each customer or- model, are provided here.
der’s delivery time. Although the data are highly
granular, we do not have information on product Logistics Rating. Logistics rating (the dependent vari-
categories (e.g., book or electronics), customer demo- able) is the customer’s rating for their experience with
graphics (e.g., location of the customer, gender, or the logistics service for their order. This variable uses
age), seller or carrier region, or marketplaces in the a Likert scale with values ranging from 1 (low quality)
data set provided by the MSOM data-driven research to 5 (high quality). We find that 62.2% of the customer
challenge. Hence, our research cannot address ques- orders have a “no response” for the rating in Table 2.
tions related to heterogeneity arising from product Excluding these observations, which are substantial in
category, customer type, region, or marketplaces. number, can potentially lead to biased estimates. Hence,
Table 1 provides an overview of the data set. we provide an econometric specification which models
A total of 13,951,407 unique customers placed “no response” option, as described later.
15,848,263 orders during April 2017. We focused on the The independent variables are as follows.
top 14 carrier companies, which handled 95% of the to-
tal volume for our analysis. The largest carrier compa- Delivery Time. Delivery time is the time duration
ny, identified as 184 in the data, delivered more than from the moment a customer placed the order to the
25% of total customer orders. We found that each cus- time (s)he received the package. We measure the unit
tomer order had a median paid amount of 44.6 yuan. of this variable in “days,” which is the same unit of
Table 2 lists the distribution of the customer’s re-
measure as the customer’s requested promise speed.
quested promise speed, delivery time, and logistics
rating. We find a significant number (91.8%) of orders
Pay Norm. Pay norm is the scaled measure of the total
had no promise speed, meaning these orders have no
amount paid for each customer order. This variable is
time restriction on when the customer can expect to
scaled to a mean of zero and standard deviation of
receive the package. The remaining 8.2% of the orders
one. We use this scaled measure rather than the actual
account for two-day and one-day promise speed re-
value for two reasons: (i) the variation in the total
quests. Around 95% of the orders are delivered within
amount paid in the data are relatively large (mini-
five days, and most orders are delivered in three days.
mum 0.01 yuan and maximum 1,215,146 yuan) and
We find that 62.2% of customer orders contained no
logistics service rating in our data. The remaining (ii) to prevent having a small number for the coeffi-
37.8% of orders with a logistics rating have a J-shaped cient in the regression results.
distribution for the ratings, which is commonly ob-
served on online platforms (Hu et al. 2009). Cainiao. Cainiao is a binary variable that takes the val-
ue of one if Cainiao managed the customer order and
4. Impact of Delivery Time on a value of zero if managed by the independent seller.
Logistics Rating Controls. We use several controls in our model since
In this section, we analyze the first phase of our re-
the underlying customer’s utility to post a rating (or
search question: What is the impact of delivery time on
no rating), and the ratings they provide from logistics
the logistics rating provided by the customer for the
service is likely to depend on additional variables.
seller from whom they purchased the product? (Q1a).
Our controls include time-invariant variables such as
the seller from whom the product was purchased, the
Table 1. Cainiao Network Data: April 2017
carrier company that delivered the order, and time ef-
Variable Value fects such as the week of the month, day of the week,
hour of the day, and holidays. Each seller and the
Customer orders 15,848,263
Unique customers 13,951,407 carrier company may follow different inventory man-
Unique sellers 345 agement and shipment policies, which are not directly
Unique items 139,354 observed in the data. Hence, we control for these
Total revenue 1.878 Billion Chinese Yuan (≈ $0.270 Billion) omitted variables through the seller, carrier company
Carrier companies 14
fixed effects, and time effects.
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 833

Table 2. Distribution of Promise Speed, Delivery Time, and Logistics Rating

Promise speed Delivery time Logistics rating

Status Count Proportion Value (days) Count Proportion Status Count Proportion

No-promise 14,546,267 91.8% 1 131,532 0.8% No response 9,849,239 62.2%

Two-day 1,167,622 7.4% 2 2,819,844 17.8% 1 90,982 0.6%
One-day 134,374 0.8% 3 5,597,409 35.3% 2 31,902 0.2%
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

4 4,548,837 28.7% 3 134,372 0.8%

5 1,825,257 11.5% 4 356,404 2.2%
≥6 925,384 5.8% 5 5,385,364 34.0%

4.2. Model and Results Yit is equal to r is the probability that X it · β +
2it falls
We explain the underlying data generating process of between Kr−1 and Kr:
observed logistics ratings. The dependent variable, lo- Pr (Yit  r)  Pr (Kr−1 < X it · β +
2it ≤ Kr ): (2)
gistics rating, is a discrete ordered outcome ranging
from 1 to 5 with 1 being a low rating and 5 being a Yit is the ordinal outcome for order i placed at time t,
high rating. In our data set, we find 62.2% of the cus- whereas the vector X it comprises all covariates in the
tomer orders with rating status “no response”. Hence, model. γ, β represent the coefficients of the selection
our data sample is likely to suffer from the selection model and ordinal outcome model, respectively. The
on unobservables problem (De Luca and Perotti 2011) variable r takes values of one, two, three, four, and
because the errors that determine whether a customer five. K0 to K5 are cut-off points on the underlying cus-
posts (or does not post) a rating are potentially corre- tomer latent response curve with K0  −∞ and
lated with errors that determine the value of a rating. K5  ∞.
1it ,
2it are the idiosyncratic random error
Next, we describe a model that addresses this issue. terms in respective models and are assumed to follow
There are two dependent variables in our data gen- a bivariate normal distribution with a mean of zero
erating process. First, there is a binary outcome, sit and a correlation coefficient of ρ.
that indicates whether customer i in the sample has The Heckman ordered model, although new to the
provided a rating. Second, there is an ordinal out- operations management field, is a commonly used
come, which is the actual rating provided by customer model in marketing and management literature uti-
i, Yit, conditional on customer i posting a rating. Both lized to understand the drivers of customer-provided
outcomes are categorical. We model both outcomes ratings. The selection model is often referred to as an
jointly as the errors in their data generating process incidence model (Moe and Schweidel 2012, Lee et al.
are likely to be correlated because of sample selection. 2015, Karaman 2020) or propensity model (Ho et al.
The outcomes are modeled as a linear combination of 2017b). The ordered outcome model is referred to as
covariates relative to specific cutoff points that parti- evaluation model (Moe and Schweidel 2012, Ho et al.
2017b) or rating response model (Lee et al. 2015). A
tion the real line. With idiosyncratic errors that are
distinctive feature of the Heckman models is that they
normally distributed, the econometric specification is
account for and help resolve the sample-induced en-
a combination of the Probit selection model and the
dogeneity resulting from non-random sample selec-
Ordered Probit rating model, referred to as the Heck-
tion such as truncated and censored samples (Certo
man Ordered Probit Regression model (De Luca and
et al. 2016). However, this model’s primary limitation
Perotti 2011). An important feature of this model is
is that it does not address the endogeneity2 resulting
that it does not discard no-rating response observa-
from independent variables (Certo et al. 2016).
tions and models both the customer’s rating decision
We ran the Heckman ordered probit regression
and the ratings that they provide in the data generat-
model on each subset of promise speed orders (no-
ing process. Although the Bivariate probit model has
promise, two-day promise, and one-day promise) sep-
been used in prior operations literature (Kim et al.
arately. Table 3 provides the regression results of all
2015), we are not aware of the application of the Heck- the models. Column (1) lists all the covariates, con-
man ordered probit model in an operations context. trols, and model statistics. Across all the models, the
The model specification is as follows. reference level of the delivery time variable is one
The selection model is given by day. Columns (2) and (3) list the probit selection and

ordered probit regression results on No-promise
1 if (X it · γ +
1it > 0)
sit  : (1) speed customer order data. From column (2), we find
0 otherwise
that the coefficient of delivery time increases with lon-
The ordinal outcome model for the logistics rating, Yit, ger delivery time. This result suggests that the likeli-
is as follows. The probability that the ordinal outcome hood of a no-promise speed customer posting a rating
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
834 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

increases with an increase in delivery time. From more likely to post a rating when the magnitude of
column (3) (ordinal outcome model), we find that the disconfirmation (s)he encounters is large. The mea-
coefficient of delivery time decreases with longer de- sure for disconfirmation in Ho et al. (2017b) is derived
livery time. This result suggests that ratings provided from the distribution of ratings before purchase. We
by no-promise speed customers decrease with an in- find a similar result in the context of no-promise
crease in delivery time. A no-promise speed customer speed customers but using an accurate measure for
does not have a set deadline for order delivery. The discomfort. Hence, our results strengthen the underly-
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

seller or carrier provider is allowed to deliver the ing theory and provide unbiased results on the impact
package for these orders at their convenience. Surpris- of discomfort on customer satisfaction. Drawing
ingly, we find that the chance that no-promise speed on parallels from expectation-disconfirmation theory
customers (non-priority customers) post a logistics and Ho et al. (2017b), we surprisingly find that a
rating is higher, and the value of logistics rating is no-promise speed customer has an average expecta-
lower with delivery time greater than two days. tion of two-day delivery time, although they are not
Expectation-disconfirmation theory (Oliver 1977, promised any specific delivery time.
1980) has been extensively used to study customer sat- There are two possible explanations for the two-day
isfaction, as discussed in the literature review section. delivery time being a hypothetical baseline expecta-
In both columns (2) and (3), we find the coefficient of tion of logistics service for no-promise speed custom-
two-day delivery time to be insignificant. Therefore, if ers. First, a two-day delivery has become a logistics
we consider two-day delivery time as a no-promise service quality norm for most online retailers (Crook
speed customer’s hypothetical baseline expectation of 2016, Liptak 2019, Del Rey 2020). As a result, custom-
logistics service, orders delivered in three days, four ers will likely be dissatisfied with the service and pro-
days, five days, and later denote increasing order of vide a lower rating for delivery time beyond two
discomfort. As a result, our results can be inferred as days. Second, the alternate explanation is based on
follows. A no-promise speed customer is more likely customers anchoring their delivery expectations at
to post a rating and provide a lower rating with in- two days or a little longer than two days. The anchor-
creasing discomfort. Analyzing data on complete ing effect theory suggests that people make predic-
purchasing and rating activities on an e-commerce tions about uncertain events by starting from an initial
website, Ho et al. (2017b) finds that an individual is value and then continually adjusting to reach the

Table 3. Heckman Ordered Probit Regression Results

No-promise speed Two-day promise speed One-day promise speed

Probit Ordered Probit Ordered Probit Ordered

Variable selection Probit Variable Selection Probit Variable selection Probit
(1) (2) (3) (4) (5) (6) (7) (8) (9)

Delivery 2 days 0.034 −0.048 2 days 0.055 −0.170 2 days −0.051 −0.350***
time (0.019) (0.047) (0.039) (0.192) (0.037) (0.035)
3 days 0.059** −0.134** 3 days 0.041 −0.539** ≥ 3 days −0.041 −0.350***
(0.019) (0.046) (0.039) (0.196) (0.035) (0.105)
4 days 0.086*** −0.210*** ≥ 4 days −0.028 −0.796***
(0.019) (0.046) (0.039) (0.201)
5 days 0.105*** −0.295***
(0.019) (0.046)
≥ 6 days 0.074*** −0.456***
(0.019) (0.048)
Pay norm 0.002 0.049*** 0.012* 0.031** 0.007** 0.015
(0.002) (0.009) (0.005) (0.009) (0.003) (0.016)
Cainiao 0.007 −0.002 −0.046 −0.020 −0.009 0.018
(0.016) (0.014) (0.029) (0.018) (0.003) (0.039)
ρ 0.000 −0.005 −0.022
(0.055) (0.212) (0.479)
Controls Seller, Carrier Yes Yes Yes Yes Yes Yes
Week, Day, Hour Yes Yes Yes Yes Yes Yes
Holidays Yes Yes Yes Yes Yes Yes
Observations 14,546,267 5,507,349 1,167,622 440,914 134,374 50,761
Log Likelihood −12,113,335.0 −909,463.4 −102,589.3
Note. Standard errors clustered at seller ∗ carrier level.
*p < 0.05; **p < 0.01; ***p < 0.001.
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 835

final value (Tversky and Kahneman 1974, Epley and 0.0135. The average marginal effects for Two-day prom-
Gilovich 2001). In examining the impact of delivery ise speed and One-day promise speed data are inter-
performance on logistics rating, we primarily focus on preted along similar lines. These estimates show that
customer’s response to their focal order delivery reducing delivery time reduces the chance of a low rat-
performance. Our study does not include or model ing while increases the chance of receiving the highest
the customer’s responses based on their recent and rating for logistics service.
past service experience due to a limited number of The regression results and marginal effects provide
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

such transactions in our data. Hence, an incoming sufficient empirical evidence on the significant rela-
no-promise speed customer is likely to anchor the ex- tionship between delivery performance and logistics
pectation of their focal order delivery performance rating. To support the causal claims of our model, we
based on the available shipping options of certainty conducted extensive analysis to address issues such as
on the platform. The Tmall/Cainiao platform pro- endogeneity and performed robustness checks. We re-
vided only two fast shipping choices, one-day and port these results in Section 7 of the paper and Section
two-day, during April 2017 (our data period). A A of the online appendix. The results and insights of
no-promise speed customer neither pays an additional this analysis strongly support the causal inference
price for logistics service nor sets the deadline for or- drawn from our model.
der delivery. Hence, we believe these customers are
most likely to anchor their delivery expectation at the 5. Impact of Logistics Rating on
slowest shipping option on the platform (two days) or
a little longer (three days or four days, etc.). Although
Customer Purchase Probability
In this section, we analyze the second phase of our
the possibility of heterogeneous anchoring expecta-
main research question: What is the impact of im-
tions exists across customers, our results show that
proved logistics rating for a seller on the customer
no-promise speed customers anchor their expectation
purchase probability and the sales of third-party sell-
at the slowest guaranteed shipping option of two
days available on the platform. ers on the platform? (Q1b).
The insights from both the Two-day promise speed We use the choice modeling framework to examine
and One-day promise speed models are as follows. the impact of logistics ratings. In the traditional choice
From columns (5) and (8), we find delivery time has modeling literature (Train 2009), customers face the
no impact on these customers’ likelihood of posting a purchase decision of a product from a finite number
rating. However, from columns (6) and (9), we find of channels (C). The customer with an intent to pur-
coefficients of delivery time decrease with longer de- chase compares the product from each channel based
livery time beyond their anticipated delivery speed. on attributes such as price, quality, channel prefer-
This result suggests that ratings provided by these ence, and additional variables. The customer receives
customers decrease with an increase in delivery time the product’s indirect utility, depending on the value
beyond their expected delivery date. We do not find and weight attached to each attribute. For example, a
strong statistical significance of the channel (Cainiao budget-constrained customer receives higher utility
or seller) variable across the selection and ordered for a lower-priced product and a lower utility for a
outcome models. This result indicates that the channel higher price. The net utility is assumed to be a weight-
of delivery does not impact logistics ratings. ed linear function of all product attributes and un-
The Heckman ordered probit regression is a nonlin- observable factors. The theory of utility-maximizing
ear model whose coefficient estimates cannot be di- behavior predicts that the customer compares the
rectly interpreted as marginal effects. From columns product’s net utility from different channels and
(2), (3), (5), and (6) in Table 3, we find that the coefficient makes a decision to purchase from one of the channels
of two-day delivery time is statistically insignificant. As or not to purchase at all (typically referred to as the
a result, we combine the reference level one-day and “outside option”). Each channel’s choice probabilities
two-day delivery time category into a single category or market shares derived under the standard MNL
for these customers to generate marginal effects. Table 4 framework depend on the difference between the utili-
lists the average marginal effects of the delivery time ties. The outside option’s utility is often set to a refer-
for all the models. The estimates from column (2) are in- ence value of zero because the product is assumed to be
ferred as follows. If a three-day delivered order is deliv- unavailable for purchase beyond the C channels. This
ered in one day or two days, it decreases the average assumption does not hold in the context of our study.
probability of a no-promise customer posting a rating We are constrained with the data on products, their
by 0.0094. For the same customers, the average proba- attributes, and sales information only from sellers op-
bility of posting a rating of 1, 2, 3, or 4 decreases by erating on a single platform, that is, Tmall. We do not
0.0026, 0.0008, 0.0032, and 0.0069, respectively, whereas have data on similar products sold from sellers on oth-
the probability of posting a rating of 5 increases by er competing online platforms (e.g., JD.com) or brick-
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
836 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

Table 4. Average Marginal Effects of Delivery Time

No-promise speed Two-day promise speed One-day promise speed

3 days 4 days 5 days ≥ 6 days 3 days ≥ 4 days 2 days ≥ 3 days

→ → → → → → → →
Response 1/2 day(s) 1/2 day(s) 1/2 day(s) 1/2 day(s) 1/2 day(s) 1/2 day(s) 1 day 1 day
Model (1) (2) (3) (4) (5) (6) (7) (8) (9)
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

Probit selection Pr(1(Rating > 0)) –0.0094 –0.0196 –0.0269 –0.0150 0.0053 0.0306 0.0194 0.0139
Ordered Probit Pr(Rating  1) –0.0026 –0.0054 –0.0091 –0.0182 –0.0144 –0.0330 –0.0125 –0.0125
Pr(Rating  2) –0.0008 –0.0017 –0.0027 –0.0052 –0.0032 –0.0067 –0.0021 –0.0021
Pr(Rating  3) –0.0032 –0.0064 –0.0103 –0.0190 –0.0120 –0.0240 –0.0090 –0.0090
Pr(Rating  4) –0.0069 –0.0136 –0.0214 –0.0375 –0.0307 –0.0570 –0.0262 –0.0262
Pr(Rating  5) 0.0135 0.0271 0.0435 0.0799 0.0603 0.1207 0.0498 0.0498
Note. Bold values indicate significance less than 1%.

and-mortar stores. The possibility that customers can The term X ijt is a vector of different covariates of the
access similar products from these other channels item sold by seller i. In our context, X ijt includes both
exists. As a result, customers can receive a non-zero time-varying covariates such as the average price of the
utility of purchase from other channels violating the item on day t, cumulative average logistics rating of
assumption of zero utility for the outside option in the seller i until day t, average order quality of the item from
standard MNL model. We use the framework in New- seller i on day t, and other covariates such as day of the
man et al. (2014), which proposes an efficient and week, week of the month, and holidays. The parameter
consistent estimation of choice-based models when αij in the utility specification is a constant that captures
one alternative is completely censored. Other channels customer’s preference to the item sold by the seller i. The
of purchase, such as competing online platforms and parameter vector β captures the effect of each covariate
brick-and-mortar stores, are entirely censored in our on the utility specification. The idiosyncratic component
context. We build on this choice model and extend
ijt captures the random customer utility component that
it to incorporate information on actual market varies between all the ns alternatives. We consider the
size through the number of unique customer visits to utility from the no-purchase alternative as
the platform and allow utility specification of the V0jt  γ + Z0jt · δ +
0jt :
no-purchase alternative to vary by time, which has not
been considered in the prior literature. The parameter γ is a nonzero constant that captures
We now develop a generic choice model for our the average utility of the no-purchase alternative.
context for an item sold by multiple sellers on the Newman et al. (2014) consider only a constant γ in the
platform. Consider an item j sold by a finite number previous specification of V0jt . We allow the utility
of sellers (ns) on the Tmall platform. We assume that specification of the no-purchase alternative to vary
customers randomly arrive at the platform following across the day of the week, week of the month, and
a Poisson process with an arrival rate of λ. A cus- holidays (Perdikaki et al. 2012, Lu et al. 2013, Fisher
tomer arriving to the platform is confronted with a et al. 2018); Z0jt includes all these time covariates. The
choice to purchase the item from one of the ns sellers, idiosyncratic component
0jt captures random custom-
or purchase from other online platforms, or brick- er utility that varies between all the unobserved alter-
and-mortar stores, or to not purchase from any of natives. For convenience, we remove the subscripts
these channels. For modeling convenience, we bun- j and t from the notation and develop purchase proba-
dle the following choice options: purchasing the bility expressions. Under the assumption of an inde-
item from other online or physical channels and the pendent and identical Gumbel distribution for each of
no purchase option into a single category and refer it the idiosyncratic components (McFadden 1984), the
as a no-purchase alternative. Hence, the mutually expression for the choice probability of purchasing
exclusive and completely exhaustive alternatives the item from seller i and no-purchase alternative are
comprise purchasing from one of the ns sellers and a given by
no-purchase alternative. For an arriving customer, eαi +Xi β
the utility of purchasing item j from seller i on day t Pi (α, β, γ, δ)  ns   γ+Z δ ;
αi +X i β
i1
e +e 0
is specified as
eγ+Z0 δ
P0 (α, β, γ, δ)  ns   γ+Z δ
αi +X i β
Vijt  αij + X ijt · β +
ijt i ∈ 1, 2, 3, : : : , ns : i1
e +e 0
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 837

The probability of purchasing the item from some sell- previously discussed in the literature (Pagan 1986,
er on the platform is given by Newman et al. 2014) to estimate the parameters. In the

ns first step, we maximize the right-end part of the log-
 s
P(α, β, γ, δ)  Pi (α, β, γ, δ)  1 − P0 (α, β, γ, δ): likelihood function, LL1 (α, β)  ni1 mi · logPi~ (α, β).
i1
LL1 is a simple multinomial logit (MNL) model of
The choice probabilities are function of the parameters choice of the item among different sellers operating
α, β, γ, δ, where α  [αi ] i ∈ 1, 2, 3, : : : , ns . Similar to on the platform conditional that a purchase is made.
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

Newman et al. (2014), we consider the choice proba- The no-purchase alternative is excluded from LL1.
bility of purchase from seller i conditional that a pur- Maximizing the log-likelihood function of MNL mod-
chase from one of ns sellers is observed as el that includes only a subset of exhaustive choices
eαi +X i β produces consistent parameter estimates (McFadden
Pi~ (α, β)  ns  αi +Xi β  : 1984). This implies that maximizing LL1 that does not
e
i1 include the no-purchase alternative generates consis-
Our data provide information on daily unique num- tent parameter estimates for α, β. It can be shown that
ber of customer visits for each item (N) to the plat- LL1 is globally concave and has a unique maximum.
form. The customer order data have a unique custom- Furthermore, maximizing LL1 (α, β) does not allow
er identifier for each purchase order and clicks, which estimation of all αi ( i ∈ 1, 2, 3, : : : , ns ) because of identi-
allows us to compute the number of unique customer fication problems. To estimate γ in the no-purchase al-
visits for each item in our data set. Let mi , m represent ternative specification, one of the αi must be set to
the number of unique customer purchases of the item zero. Matching shares property (Ben-Akiva et al. 1985,
from seller i and total number of customer purchases, Ferguson et al. 2012) of a limited subset of discrete
 s choice models such as MNL allows unique identifi-
respectively, where m  ni1 mi . Using this informa-
tion, we specify the likelihood function as follows: able estimates of αi with one of them set to reference
and by matching sample shares with purchase proba-
L(λ, α, β, γ, δ) bilities. Relying on this property, we set one observed
 
e−λ (λ)N N! alternative as the reference. Later in the second step,
 P(α, β, γ, δ)m P0 (α, β, γ, δ)N−m the utility of the no-purchase alternative is estimated
N! m!(N − m)!
  against this reference. Let the estimates of α, β after
m! ns
mi the first step estimation be α̂, β̂. In the second step,
Pi~ (α, β) : (3) we estimate parameters λ, γ, δ by maximizing the re-
m1 !m2 !: : : mns ! i1
maining part of the log-likelihood function LL2
The likelihood function has three components. The (λ, γ, δ)  −λ + N · logλ + m · logP(α̂, β̂, γ, δ) + (N − m) ·
first component refers to the probability that N unique logP0 (α̂, β̂, γ, δ). Note that α, β are replaced with α̂, β̂
customers visit the item page on the platform. The in LL2. The estimated values after the second step are
second component refers to the probability that m (out λ̂, γ̂, δ̂.
of N) customers purchase on the platform. Using inde- The platform updates the logistics rating at the sell-
pendence in the sequence of N customer arrivals, the er level rather than for each item individually. The
number of observed customer purchases follows a bi- distribution of logistics rating that the seller accumu-
nomial distribution. The final part of the likelihood lates at any point in time is reflected as the same for
function captures the probability that mi customers every item sold by the seller. As a result, we run the
purchase the item from seller i conditional that m pur- estimation for each item individually to examine the
chases are observed on the platform. The parameters effect of logistics rating on customer purchase proba-
λ, α, β, γ, δ are estimated using the data through the bility. For illustration, we choose items of a seller
maximum likelihood estimation approach. The log- identified as 358 in our data. Seller 358 sold a total of
likelihood function excluding the components which 361 items, is among the bottom five sellers with the
are not a function of parameters is given by least logistics rating on the platform, and handled the
highest volume of orders during April 2017. Table 5
LL(λ, α, β, γ, δ)  (−λ + N · logλ)
lists the estimation results for three items of seller 358.
+ (m · logP(α, β, γ, δ) + (N − m) Across the items, we find that the coefficient of aver-
age unit price is negative and statistically significant.
· logP0 (α, β, γ, δ))
  This result implies that higher the item price, less is
ns the utility for the customer to purchase the item,
+ mi · logPi~ (α, β) : (4) which is a trivial result. We find that the coefficient of
i1
average logistics rating is positive and statistically sig-
Maximizing the entire log-likelihood function is com- nificant, implying that higher the average logistics rat-
putationally intense. We use the two-step approach ing of a seller, higher is the utility for the customer to
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
838 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

purchase the item from that seller. The constant γ is 6. Logistics Performance and Sales
statistically significant and captures the average In this section, we examine our main research ques-
utility of the no-purchase alternative compared tion: What is the impact of reducing delivery time on
with an alternative whose αi is set to 0. We test the sales for a seller on the e-commerce platform? (Q1).
robustness of the causal link between this seller’s lo- We combine the results of our analysis from Sections
gistics rating and their sales for the three items des- 4 and 5 to analyze the following specific question:
ignated in Table 5 by analyzing the instrument free What is the potential improvement in daily sales if a
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

reduced-form models using Gaussian copula correc- seller could potentially deliver all their three-day de-
tion (Park and Gupta 2012) in Section B of the online livered orders in two days (i.e., reduce delivery time
appendix. by one day for all their current three-day delivered
We extended this estimation approach to all the 361 orders)?
items sold by this seller. We found that the effect of For illustration, we first focus on three-day delivered
the average logistics rating is significant for 176 items orders and seller 358 for two reasons. First, the number
(≈ 49%) and insignificant for the remaining 185 items. of three-day delivered orders is the largest in volume
To arrive at more comprehensive insights across all (makeup ≈ 27.1%) for seller 358 (this is generally true
the sellers operating on the platform, we extended our across all the other sellers). Second, seller 358 is among
choice modeling estimation analysis for all the 119,276 the bottom five sellers with the least logistics rating on
items sold by 335 sellers.3 We found that the effect of the platform and handled the highest number of orders.
average logistics rating on customer purchase proba- Hence, reducing delivery time by one day for three-day
bility is significant for ≈ 51% of the items and insignif- delivered orders is likely to impact sales of seller 358.
icant for the remaining 49%. We conducted a t test to We first quantify the impact of shipping all three-day
examine whether the price and volume sold vary sig- delivered orders in two days on the seller’s distribution
nificantly between items with and without the signifi- and expected logistics rating. Next, we estimate the in-
cance of average logistics rating on customer purchase crease in customer purchase probability and sales be-
probability. We found that the average unit price is cause of the change in expected logistics rating.
not significantly different between the two groups. During April 2017, seller 358 handled 353,299
In contrast, the average volume is significantly higher orders with all orders shipped as no-promise speed
for items where logistics ratings have a significant orders. Of the total volume, there are no ratings avail-
impact. Based on our extensive analysis, we conclude able for 204,184 orders (57.79%). For the remaining
that the impact of logistics rating on sales is statistical- 149,115 orders (42.21%), the conditional distribution
ly significant for close to half of the items on the of ratings is as follows: Rating 1 (count 4,348; propor-
platform. Furthermore, the improvement in logistics tion 2.92%), Rating 2 (1,342; 0.90%), Rating 3 (5,417;
rating is potentially more beneficial for high volume 3.63%), Rating 4 (9,530; 6.39%) and Rating 5 (128,478;
items. 86.16%). We estimate the baseline expected rating to
In summary, the results and insights from the anal- be 4.7198. Of the total 353,299 orders, the number of
ysis of Q1a and Q1b provide empirical evidence to three-day delivered orders are 95,701.
the mechanism that connects delivery performance to Utilizing the marginal effects in Table 4, we tabulate
sales through logistics ratings. change in the distribution and expected logistics

Table 5. Impact of Logistics Rating on Customer Purchase Probability

Item 220636 Item 258478 Item 183163

Variable (1) (2) (3)

Average unit price −0.446** −0.376* −0.159*

(0.114) (0.135) (0.068)
Average logistics rating 102.887*** 50.088** 147.154*
(25.313) (15.974) (48.764)
Average order quality rating −0.851 −0.046 −12.990
(2.261) (2.085) (8.261)
γ 474.328** 233.915** 627.906*
(119.824) (79.097) (211.737)
Week, day, holidays Yes Yes Yes
Observations 30 30 22
Log Likelihood −4.986 −1.764 −7.768
*p < 0.1; **p < 0.05; ***p < 0.01.
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 839

rating for this seller due to shipping their three-day sales of item j of seller i on day t under baseline ex-
delivered orders in two days. Reducing the delivery pected logistics rating is given by Sijt1  prijt1 · Nijt · pijt ,
time of the three-day delivered orders changes the where prijt1 , Nijt , pijt denote purchase probability,
likelihood of posting a rating and the value of rating unique visitors, and average price of item j, respec-
for these customers’ orders while others remain unaf- tively, on the Tmall platform on day t under current
fected. As a result, we first separate the volume of or- baseline logistics performance. Similarly, the pre-
ders into two categories: three-day and ≠ three-day dicted sales of the same item under improved logistics
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

delivered orders. Table 6 provides a detailed enumer- rating by reducing delivery time from three days to
ation of change in the distribution of ratings. Column two days is given by Sijt2  prijt2 · Nijt · pijt , where
(1) lists the category of orders by delivery time. Col- prijt2 , Nijt , pijt denote purchase probability, unique visi-
umn (2) lists the proportion of availability of ratings tors, and average price of item j, respectively, on the
for each category. For example, 57.89% of all three- Tmall platform on day t under improved logistics per-
day delivered orders have no rating, whereas the re- formance. The difference prijt2 − prijt1 represents the in-
maining 42.11% of orders have a rating. Columns (4) cremental customer purchase probability resulting
and (5) list the count and conditional distribution of from the change from baseline to improved logistics
ratings when the rating is available. Columns (6) and rating performance. The average daily sales for
(7) list the marginal effects. Table 4 shows that deliver- seller i from all their ni items, over D days under base-
ing a three-day order in two days decreases (in- line and improved logistics rating are given by
 n i 1 D n i
creases) the average probability of a customer posting Si1  D1 D t1 j1 Sijt1 , Si2  D t1 j1 Sijt2 , respective-
a rating (no rating) by 0.0094. The probability of a cus- ly. The percentage change is average daily sales for
tomer posting a rating of 1, 2, 3, and 4 decreases by seller i is given by (Si2 =Si1 − 1) · 100%. We implement
0.0026, 0.0008, 0.0032, and 0.0069, respectively, where- this computation and find that the improvement in
as that of rating 5 increases by 0.0135. The marginal ef- average daily sales for the seller 358 to be 10.3%. In
fects for ≠ three-day delivered orders are 0 as these
conclusion, we find that reducing the delivery time of
orders remain unaffected. Using the new distribution
of ratings, we estimate the expected logistics rating to all the three-day delivered orders (≈ 27.1% of its total
be 4.7265. We estimate that shipping all the three-day orders) for the seller 358 to two days results in an in-
delivered orders (≈ 27.1% of its total orders) of seller crease of average daily sales for this seller by 10.3%.
358 in two days results in an increase in expected We repeated the entire process of estimating the
logistics rating from 4.7198 to 4.7265. impact of shipping each seller’s three-day delivered
We now estimate the increase in customer purchase orders in two days on sales to derive general insights
probability and sales because of the change in ex- for all sellers. We found that reducing the delivery
pected logistics rating. We provide the general expres- time of all three-day delivered orders (that makeup
sions for the increment in sales for a seller i. Let j, ni ≈ 35% of the total orders across all sellers) to two
represent the item identifier and the number of items days improves the average daily seller sales by 13.3%
sold by seller i, where j ∈ 1, 2, 3, : : : , ni . The predicted on the platform.

Table 6. Impact of Shipping Three-Day Delivered Orders in Two Days on the Distribution of Logistics Rating

Baseline Marginal effects Improved

Delivery Percentage Percentage Percentage Percentage Expected

time ratings (s) Rating (r) Count rating (pr) Δs Δpr ratings (s) rating (pr) rating
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

 three 57.89% --- 55,399 --- 0.94% 58.83% ---

days 1 979 2.43% −0.26% 2.17%
2 295 0.73% −0.08% 0.65%
42.11% 3 1,278 3.17% −0.94% −0.32% 41.17% 2.85%
4 2,553 6.33% −0.69% 5.64%
5 35,197 87.33% 1.35% 88.68%
≠ three 57.76% --- 148,785 --- 0.00% 57.76% ---
days 1 3,369 3.10% 0.00% 3.10%
2 1,047 0.96% 0.00% 0.96%
42.24% 3 4,139 3.80% 0.00% 0.00% 42.24% 3.80%
4 6,977 6.41% 0.00% 6.41%
5 93,281 85.73% 0.00% 85.73%
4.7265
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
840 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

To derive further insights, we summarize the results 7.1. Selection on Unobservables

on the improvement in sales by item volume group The selection on unobservable problems arises
categories and item price group categories. We first when the unit’s assignment to the treatment group is
sort the sellers in the ascending order of % three-day because of variables that are not observed and affect
volume orders handled by the seller. The top 25% and the outcome. The general quasi-experimental methods
the bottom 25% of the sellers are labeled into Low and used to estimate the causal effect of treatment
High % three-day volume group categories, respec- under the selection of unobservables problem are re-
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

tively, whereas the remaining are labeled as the Medi- gression discontinuity design (RDD) and difference-
um group category. Table 7 lists the average change in in-differences (DID) method (Angrist and Pischke
percentage sales by improving logistics performance 2008). We use the RDD approach as DID requires ob-
for each % three-day volume group category. We find servations during pretreatment and posttreatment pe-
that reducing delivery time for all three-day delivered riods for both the treatment and control groups. Such
orders (that makeup ≈ 28.9%–39.2% of the total orders) kind of data generating process is not viable in our
to two days improves the average daily sales by context. To use RDD, we need to identify a threshold
9.0%–22.5% on the platform. The impact is most signif- value of a continuous variable (also referred to as forc-
icant for sellers in the High % three-day volume group ing variable) that differentiates between the treatment
category. and control group observations. The observations
Next, we summarize the results on the improvement around the threshold value of the forcing variable are
in sales by improving logistics performance by item mostly similar. The difference in output values be-
price group. We sort the items sold on the platform in tween treatment and control observations accounting
the ascending order of prices. The top 25% and the bot- for any observable variables provides the causal esti-
tom 25% of the items are labeled as Low and High price mate of the treatment effect.
item group categories, respectively, whereas the re- In our context, we consider the time of order arrival
maining are labeled as Medium price groups. Table as the forcing variable and the threshold value of 12
a.m. (midnight) that identifies the treatment and con-
8 lists the average change in percentage sales for each
trol group orders. We focus on the no-promise speed
item price group. On average, we find that the improve-
orders that the platform received in the one-hour in-
ment in sales by decreasing delivery time by one day is
terval before and after midnight. For example, consid-
most prominent for the high-priced item category.
er the three-day delivered orders that arrived from
In addition to these results, we analyzed the sales
11:00 p.m. to 11:59 p.m. on the current day and the
improvement for each seller-item combination by the
two-day delivered orders that arrived during 12 a.m.
item price and % three-day volume seller group and
to 1 a.m. on the next day. Drawing parallels from the
across different sellers grouped by average delivery
RDD framework, if the two-day delivered orders are
time. We report these results in Section C of the online
considered in the control group, the three-day deliv-
appendix.
ered orders can be considered in the treatment group.
These orders around the threshold of 12 a.m. (mid-
7. Robustness Analysis night) are not significantly different, given the time in-
The first phase of the main research question (Q1) that stant orders were received after controlling for other
we examine is as follows: What is the impact of deliv- observable covariates. Hence, the difference in logis-
ery time on the logistics rating provided by customers tics rating between these two-day and three-day deliv-
for sellers from whom they purchased the product? ered orders provides the causal estimate of the change
(Q1a). We analyze the robustness of Q1a results be- in logistics rating because of an increase in an addi-
cause of selection on unobservables endogeneity in tional day of delivery time. The model specification is
this section. The analysis and the results of remaining as follows. Let Di be the indicator that is one if the or-
endogeneity issues, selection on observables, simulta- der is part of the treatment group and otherwise zero:
neity, and omitted variable bias, are reported in Di  1[Fi < c], Fi is the time instant the order was re-
Section A of the online appendix. ceived, and c is 12 a.m. We run the following ordered

Table 7. Improvement in Sales for Sellers Grouped by %

Three-Day Volume Table 8. Improvement in Sales for Items Grouped by Price

% Three-day order % Change in daily Unit price Percentage change in

Seller group volume (mean) sales (mean) Item price group (Yuan) (mean) daily sales (mean)

Low 28.9% 9.0% Low 6.44 13.6%

Medium 34.8% 10.8% Medium 26.59 22.0%
High 39.2% 22.5% High 287.78 28.2%
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 841

Table 9. RDD Results: One-Hour Window Around 12 a.m. (Midnight)

Control 2-day 3-day 4-day

Treatment 3-day 4-day 5-day
Variable (1) (2) (3)

D −0.1560* −0.1760** −0.1160

(0.0673) (0.0577) (0.0834)
F–c −0.0008 −0.0017*** −0.0014*
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

(0.0006) (0.0004) (0.0006)

D · (F − c) −0.0003 −0.0014** −0.0013
(0.0007) (0.0005) (0.0007)
Pay 0.0000*** 0.0001*** 0.0002***
(0.0000) (0.0000) (0.0000)
Cainiao 0.0653*** 0.0289* −0.0339
(0.0132) (0.0121) (0.0222)
Controls Seller, carrier Yes Yes Yes
Week, day, hour, holidays Yes Yes Yes
Observations 83,172 166,047 74,972
Log Likelihood −33,090.3 −74,125.3 −37,144.7
Note. Standard errors clustered at seller ∗ carrier level.
*p < 0.05; **p < 0.01; ***p < 0.001.

probit model results cannot be generalized to the entire data set.

Hence, to compare the main model (Heckman ordered
Pr (Yi  r)  Pr (Kr−1 < (α + β · Di + γ1 · (Fi − c)
probit) and RDD estimates (Table 9), we analyze the
+ γ2 · Di · (Fi − c) + X i · δ +
i ) ≤ Kr ): Heckman ordered probit model on the limited num-
Yi is the ordinal outcome for the order i. The variable r ber of observations used for the RDD model. Table 10
takes values of one, two, three, four, and five. K0 to K5 summarizes the comparison. Column (2) lists the re-
are cutoff points on the underlying customer latent re- sults of the Heckman ordered probit model employed
sponse curve with K0  −∞ and K5  ∞. The term
i is on the RDD data with two-day delivery time orders
the idiosyncratic random error term assumed to fol- (control group) vs. three-day delivery time orders
low normal distribution. The coefficient of Di, β pro- (treatment group). Column (3) captures the RDD esti-
vides the causal effect. It is unnecessary to include mates (column (1) from Table 9). A comparison of col-
other covariates in the regression, even if they are es- umns (2) and (3) shows that the three-day delivery
sential in the selection criterion. However, including time coefficient is significant and similar in magnitude
available covariates can help reduce any small-sample for both models. Similarly, columns (4) and (5) com-
bias (Imbens and Lemieux 2008). The term X i includes pare the estimates of the two models for three-day
all observable covariates such as order pay, channel, versus four-day delivery time orders. We find the esti-
seller, carrier, and time fixed effects. mates from the causal RDD method and the Heckman
Table 9 lists the RDD results. Column (1) lists the re- model are very close in all cases, and the magnitude
sults of three-day delivered treatment group orders of the bias of the Heckman model is less than 5%. This
against two-day delivered control group orders. We eases any causal concerns of the delivery time effect
find that the coefficient of D in column (1) is negative on logistics rating for no-promise speed orders in the
and statistically significant. This result indicates that Heckman model.
the logistics rating provided by three-day delivered
no-promise speed customers is lower than two-day 8. Conclusion and Managerial Implications
delivered no-promise speed customers. Similarly, in Online retail is one of the fast-growing business in
columns (2) and (3), we compare three-day versus e-commerce. Online retailers provide several advan-
four-day delivered orders and four-day versus five- tages: avoiding travel to physical stores, accessing a
day delivered orders received around one-hour inter- wide variation of products via a digital channel,
vals from midnight. To conclude, we note that if the smooth one-click purchase transaction, and many
endogeneity in our data generating process is because more. Customer experience plays a cardinal role in
of selection on unobservables, our results from RDD creating a monetary transaction for sellers on the
provide the causal effects of delivery time on logistics platform. Increasing customer traffic to the platform is
rating for no-promise speed customer orders. the key to the online retail business’s survival and
The RDD model, although causal, generates a local growth. These platforms can increase customer traffic
estimate because of the selection of no-promise orders either by investing efforts to retain and make the exist-
observed exclusively around midnight, and, hence, its ing customer shop more or attract new customers
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
842 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

Table 10. Ordered RDD Results: No-Promise Speed Orders

2 days vs. 3 days 3 days vs. 4 days 4 days vs. 5 days

Ordered Ordered Ordered

Variable Probit RDD Probit RDD Probit RDD
(1) (2) (3) (4) (5) (6) (7)

Delivery time 3 days −0.149* −0.156*

Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

(0.065) (0.067)
4 days −0.163** −0.176**
(0.058) (0.058)
5 days −0.101 −0.116
(0.083) (0.083)
Controls Pay norm, Cainiao Yes Yes Yes Yes Yes Yes
Seller, carrier Yes Yes Yes Yes Yes Yes
Week, day, hour, holidays Yes Yes Yes Yes Yes Yes
N 83,172 83,172 166,047 166,047 74,972 74,972
LL −33,090.3 −33,090.2 −74,129.5 −74,125.3 −37,146.4 −37,144.7
Note. Standard errors clustered at seller ∗ carrier level.
*p < 0.1; **p < 0.05; ***p < 0.01.

through word-of-mouth. A means to achieve this for recommendations. First, our study finds that reducing
sellers is by selling products with lower prices, superi- the delivery time of all three-day delivered orders
or quality, and provide enriching service experience (that makeup ≈ 35% of the total orders) to two days
compared with their competitors. With limited varia- improves the average daily sales by 13.3% on the plat-
tions in prices and product quality across competing form. This result has an immense practical significance
platforms, service experience can be a differentiating for online retailers, as it quantifies the potential benefit
factor in attaining a market share. As a result, online of investing in improved logistics performance. Be-
retailers typically provide premium membership, sides, the result emphasizes that delivery time perfor-
which promises fast delivery services such as same- mance and logistics ratings, which measure service
day or next-day deliveries, hoping that customers are quality, are essential drivers of a customer’s likelihood
impressed by an excellent service experience, provide of purchase and, hence, sales on e-commerce platforms.
useful feedback to the platform, and return to shop in Hence, we recommend that e-commerce platforms and
the future. Faster deliveries, although convenient to sellers should pay attention to logistics performance
customers, results in high costs to online retailers. The quality, in addition to product quality in driving traffic
additional costs range from starting new fulfillment and sales.
centers to air-cargo operations. As a result, online re- Second, we find that longer delivery time increases
tailers need to examine if the investment in fast ship- the likelihood of posting a logistics rating and results in
ping infrastructure has a payoff in increasing sales. lower logistics rating by non-priority customers (no-
The mechanism of improved financial outcomes promise speed customers in our context) for delivery
from superior logistics performance can be a result of a time beyond two days. This result is surprising because
customer’s response to (i) their own recent or past ser- it suggests that while customers with no promise speed
vice experience or (ii) to accumulated feedback from do not have an explicit delivery deadline, they seem
previous customers’ experience on the common plat- to assume a two-day delivery deadline implicitly.
form (word-of-mouth). This paper focuses on the sec- Hence, they are disappointed by deliveries that take
ond mechanism (word-of-mouth effect through logistics longer than two days. Thus, no-promise speed cus-
ratings) by providing an empirical validation of this tomers seem to have been conditioned to expect de-
mechanism. We analyze the mechanism by examining liveries in two days. Hence, e-commerce platforms
the following two research questions individually. First, should not take no-promise speed customers for
what is the impact of delivery time on the logistics rat- granted and should pay attention to these customers’
ing provided by a customer? (Q1a). Second, what is the delivery performance.
impact of improved logistics ratings on customer pur- Third, for priority customers (two-day and one-day
chase probability and sales for the seller? (Q1b). We promise speed customers), we find that longer deliv-
then combine our results from Q1a and Q1b to answer ery time results in lower logistics ratings for delivery
our main research question Q1. time beyond their anticipated delivery date but has no
Next, we summarize the principal results and in- impact on their likelihood of posting a logistics rating.
sights from our study and provide relevant managerial In summary, for any customer type, we conclude that
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 843

shortening delivery time results in a higher rating because of improved delivery performance can offset
conditional that the customer provides a rating. Be- the increase in costs of investing in additional infra-
cause customers post a higher rating upon observing structure for faster shipping. By quantifying the im-
superior delivery performance, increasing the likeli- pact of delivery time performance on sales, our study
hood of customer posting a rating is a key to achieve a provides a method to assess the benefits of faster de-
significant jump in logistics rating for the seller and liveries. Our insights are relevant to independent sell-
on the platform. As a result, we suggest that the plat- ers and e-commerce platform managers who aim to
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

form or sellers reach out to customers and encourage improve long-term online customer traffic and sales.
them to participate in the rating process, especially Like every research study, our study also has its limi-
when they experience superior delivery performance. tations. For example, our study does not model the stra-
Fourth, our results suggest that there is no addition- tegic interaction among the sellers on the platform. In
al incentive for delivering orders earlier than the other words, our study does not provide answers to
anticipated delivery date for priority customers. For questions such as how should other sellers modify their
example, we find that delivering two-day promise delivery speed options when a focal seller provides a
speed customer orders in one day does not provide faster shipping speed. For example, should other sellers
additional value either in the increased likelihood of offer a “same day delivery” promise speed when the fo-
posting a rating or a higher rating. As a result, if the cal seller offers the same improved shipping speed op-
goal is to improve logistics rating on the platform, we tion? These issues are of importance for future research
suggest that managers need to alter their shipment but outside the scope of our current study because of
policy such that priority orders are delivered close to limitations of our data set. In conclusion, our model
their anticipated delivery date and allocate any re- and estimates apply to settings with limited strategic in-
maining capacity to delivering no-promise speed or- teraction between sellers operating on and across plat-
ders earlier. For example, consider a simple scenario forms. We leave the exploration of such further ques-
where a seller receives 10 no-promise speed and 10 tions for future research.
two-day promise orders on a given day. Let us as-
sume that the seller knows that all the no-promise or- Acknowledgments
ders can be delivered to the customers in two days, The authors thank the 2018 MSOM Data-driven Research
and all the two-day promise orders can be delivered Challenge committee for making the Cainiao Network
in one day. Assuming the seller has a capacity of ship- (Alibaba's logistics arm) data available to INFORMS mem-
ping a maximum of 10 orders per day, a policy of bers. This initiative and competition have led to this re-
shipping all no-promise speed orders on the current search paper.
day and two-day promise orders on the next day is
better than a policy that ships all two-day orders be- Endnotes
1
fore no-promise orders. This is because both In examining the impact of logistics ratings on customer purchase
no-promise speed and two-day promise speed orders probability, we primarily consider focal customer’s responses to
feedback from accumulated ratings by other customers or word-of-
will be delivered in two days with the first policy,
mouth (Dichter 1966, Hennig-Thurau et al. 2004). We exclude the
which is better than delivering two-day promise customer’s responses based on their recent and past service experi-
speed orders in one day and no-promise speed orders ence due to a limited number of such transactions in our data. We
in three days. This example suggests that always pri- gathered the count of orders for each unique customer and seller com-
oritizing orders with a promise speed over orders bination in our data. We found the median and 95th percentile of the
count to be one. These statistics indicate that most Cainiao customers
with no promise speed may be suboptimal.
have a single transaction with any seller in our data, which points to
Fifth, our analysis provides a specific recommenda- limited prior experience for any customer-seller pair.
tion on which orders (three-day delivered orders) 2
We analyzed and reported the results on multiple forms of endo-
retailers should focus on for delivery time improve- geneity emanating from independent variables such as selection
ment. Our results show that sellers can achieve the bias (selection on observables and unobservables), simultaneity,
most significant improvement in sales by reducing and omitted variable bias (Ho et al. 2017a) in Section 7 of the paper
the delivery time of no-promise speed three-day de- and Section A of the online appendix. This analysis demonstrates
and strengthens the evidence of the causal relationship between de-
livered orders by one day, which is better than other
livery performance and customer-provided logistics ratings.
choices, such as delivering four-day delivered orders 3
We excluded 10 sellers as the sales of their items are observed for
in three days. This is because a large chunk of orders less than five days in April 2017.
(>90%) on the Tmall platform are no-promise speed
orders, and the mode of the delivery time histogram
for no-promise speed orders is three days (≈ 35%). References
Allon G, Federgruen A, Pierson M (2011) How much is a reduction
Last, our study also provides a strong methodologi- of your customers’ wait worth? An empirical study of the fast-
cal framework that managers can potentially use and food drive-thru industry based on structural estimation meth-
implement to assess if the future increase in sales ods. Manufacturing Service Oper. Management 13(4):489–507.
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
844 Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS

Anderson EW, Sullivan MW (1993) The antecedents and con- Hennig-Thurau T, Gwinner KP, Walsh G, Gremler DD (2004) Elec-
sequences of customer satisfaction for firms. Marketing Sci. tronic word-of-mouth via consumer-opinion platforms: What
12(2):125–143. motivates consumers to articulate themselves on the Internet? J.
Angrist JD, Pischke JS (2008) Mostly Harmless Econometrics: An Em- Interactive Marketing 18(1):38–52.
piricist’s Companion (Princeton University Press, Princeton, NJ). Herrera S, Qian V (2019) How Amazon’s Shipping Empire Is
Ben-Akiva ME, Lerman SR, Lerman SR (1985) Discrete Choice Analy- Challenging UPS and FedEx. Accessed December 31, 2021,
sis: Theory and Application to Travel Demand, vol. 9 (MIT Press, https://www.wsj.com/articles/how-amazons-shipping-empire-
Cambridge, MA). is-challenging-ups-and-fedex-11567071003.
Ho YC, Wu J, Tan Y (2017b) Disconfirmation effect on online rating
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

Berry S, Levinsohn J, Pakes A (1995) Automobile prices in market

equilibrium. Econometrica 63(4):841–890. behavior: A structural model. Inform. Systems Res. 28(3):626–642.
Bhattacherjee A (2001) Understanding information systems continu- Ho TH, Lim N, Reza S, Xia X (2017a) OM forum: Causal inference
ance: An expectation-confirmation model. Management Inform. models in operations management. Manufacturing Service Oper.
Systems Quart. 25(3):351–370. Management 19(4):509–525.
Bray RL (2020) Operational transparency: Showing when work gets Hu N, Zhang J, Pavlou PA (2009) Overcoming the J-shaped distri-
done. Manufacturing Service Oper. Management., ePub October bution of product reviews. Comm. ACM 52(10):144–147.
29, https://doi.org/10.1287/msom.2020.0899. IBISWorld (2019) E-commerce & online auctions in the US: industry
Certo ST, Busenbark JR, Woo HS, Semadeni M (2016) Sample selec- data, trends, stats. Accessed December 31, 2021, https://www.
tion bias and Heckman models in strategic management re- ibisworld.com/united-states/market-research-reports/e-commerce-
search. Strategic Management J. 37(13):2639–2657. online-auctions-industry/.
Chen Y, Xie J (2008) Online consumer review: Word-of-mouth as a Imbens GW, Lemieux T (2008) Regression discontinuity designs: A
new element of marketing communication mix. Management guide to practice. J. Econometrics 142(2):615–635.
Sci. 54(3):477–491. Karaman H (2020) Online review solicitations reduce extremity bias
Chen Y, Harper FM, Konstan J, Li SX (2010) Social comparisons and in online review distributions and increase their representative-
contributions to online communities: A field experiment on ness. Management Sci. 67(7):4420–4445.
movielens. Amer. Econom. Rev. 100(4):1358–1398. Kesavan S, Lambert S, Williams J, Pendem P (2020) Doing well by
Chevalier JA, Mayzlin D (2006) The effect of word of mouth doing good: Improving store performance with responsible
sales: Online book reviews. J. Marketing Res. 43(3):345–354. scheduling practices at the Gap, inc. Preprint, submitted XX,
Chintagunta PK, Gopinath S, Venkataraman S (2010) The effects of https://dx.doi.org/10.2139/ssrn.3731670.
online user reviews on movie box office performance: Account- Ketokivi M, McIntosh CN (2017) Addressing the endogeneity dilem-
ing for sequential rollout and aggregation across local markets. ma in operations management research: Theoretical, empirical,
Marketing Sci. 29(5):944–957. and pragmatic considerations. J. Oper. Management 52:1–14.
Clement J (2020) Amazon: third-party seller share 2020. Accessed Kim SH, Chan CW, Olivares M, Escobar G (2015) ICU admission
December 31, 2021, https://www.statista.com/statistics/ control: An empirical study of capacity allocation and its impli-
259782/third-party-seller-share-of-amazon-platform/. cation for patient outcomes. Management Sci. 61(1):19–38.
Crook J (2016) Walmart squares up against Amazon with 2-day de- Koh NS, Hu N, Clemons EK (2010) Do online reviews reflect a
livery across the US. Accessed December 31, 2021, https:// products true perceived quality? An investigation of online
techcrunch.com/2016/06/30/walmart-squares-up-against-amazon- movie reviews across cultures. EC Res. Appl. 9(5):374–385.
with-2-day-delivery-across-the-u-s. Kulkarni VG (2016) Modeling and Analysis of Stochastic Systems (CRC
Cui R, Li M, Li Q (2019) Value of high-quality logistics: Evidence Press, Boca Raton, FL).
from a clash between SF Express and Alibaba. Management Sci. Larcker DF, Rusticus TO (2010) On the use of instrumental variables
66(9):3879–3902. in accounting research. J. Accounting Econom. 49(3):186–205.
De Luca G, Perotti V (2011) Estimation of ordered response models Lee YJ, Hosanagar K, Tan Y (2015) Do I follow my friends or the
with sample selection. Stata J. 11(2):213–239. crowd? Information cascades in online movie ratings. Manage-
Del Rey J (2020) Amazon Prime now has more than 150 million ment Sci. 61(9):2241–2258.
members. Accessed December 31, 2021, https://www.vox.com/ Levine DI, Toffel MW (2010) Quality management and job qual-
recode/2020/1/30/21115859/amazon-prime-150-members-video- ity: How the ISO 9001 standard for quality management sys-
one-day-shipping. tems affects employees and employers. Management Sci. 56(6):
Dellarocas C (2006) Strategic manipulation of Internet opinion fo- 978–996.
rums: Implications for consumers and firms. Management Sci. Li X, Hitt LM (2008) Self-selection and information role of online
52(10):1577–1593. product reviews. Inform. Systems Res. 19(4):456–474.
Dichter E (1966) How word-of-mouth advertising works. Harvard Li LI, Xiao E (2013) Money talks: Rebate mechanisms in reputation
Bus. Rev. 44:147–166. system design. Management Sci., ePub ahead of print August 1,
Duflo E, Banerjee A (2017) Handbook of Field Experiments (Elsevier, http://dx.doi.org/10.2139/ssrn.1597764.
New York). Lin CL, Lee SH, Horng DJ (2011) The effects of online reviews on
Epley N, Gilovich T (2001) Putting adjustment back in the anchoring purchasing intention: The moderating role of need for cogni-
and adjustment heuristic: Differential processing of self-generated tion. Soc. Behav. Personality 39(1):71–81.
and experimenter-provided anchors. Psych. Sci. 12(5):391–396. Liptak A (2019) Target is now offering same-day delivery directly
Ferguson ME, Garrow LA, Newman JP (2012) Application of dis- through its website. Accessed December 31, 2021, https://
crete choice models to choice-based revenue management prob- www.theverge.com/2019/6/13/18677980/target-same-day-deliv
lems: A cautionary note. J. Revenue Pricing Management 11(5): ery-website-online-shopping.
536–547. Lu Y, Musalem A, Olivares M, Schilkrut A (2013) Measuring the effect of
Fisher M, Gallino S, Li J (2018) Competition-based dynamic pricing queues on customer purchases. Management Sci. 59(8):1743–1763.
in online retailing: A methodology validated with field experi- Mao W, Ming L, Rong Y, Tang CS, Zheng H (2019) Faster deliveries
ments. Management Sci. 64(6):2496–2514. and smarter order assignments for an on-demand meal delivery
Fisher ML, Gallino S, Xu JJ (2019) The value of rapid delivery in om- platform. Preprint, submitted October 23, https://dx.doi.org/10.
nichannel retailing. J. Marketing Res. 56(5):732–748. 2139/ssrn.3469015.
 Deshpande and Pendem: Logistics Performance and Sales in E-commerce Platforms
Manufacturing & Service Operations Management, 2023, vol. 25, no. 3, pp. 827–845, © 2022 INFORMS 845

Mattioli D (2019) Amazon’s profit hurt by push to speed up ship- www.statista.com/chart/18751/physical-gross-merchandise-

ping. Accessed December 31, 2021, https://www.wsj.com/ sales-on-amazon-by-type-of-seller/.
articles/amazons-third-quarter-profit-slides-26-11571949037. Roberts MR, Whited TM (2013) Endogeneity in empirical corporate
McFadden DL (1984) Econometric analysis of qualitative response finance. Handbook of the Economics of Finance, vol. 2 (Elsev-
models. Handbook of Econometrics, vol. 2 (Elsevier, New York), ier, New York), 493–572.
1395–1457. Rosenbaum PR, Rubin DB (1983) The central role of the propensity score
McKinney V, Yoon K, Zahedi FM (2002) The measurement of web- in observational studies for causal effects. Biometrika 70(1):41–55.
customer satisfaction: An expectation and disconfirmation ap- Rossi PE (2014) Even the rich can make themselves poor: A critical
proach. Inform. Systems Res. 13(3):296–315. examination of IV methods in marketing applications. Market-
Downloaded from informs.org by [103.141.127.88] on 18 May 2024, at 10:33 . For personal use only, all rights reserved.

Moe WW, Schweidel DA (2012) Online product opinions: Incidence, ing Sci. 33(5):655–672.
evaluation, and evolution. Marketing Sci. 31(3):372–386. Schlosser AE (2005) Posting vs. lurking: Communicating in a multi-
Montasell G (2020) U.S. top online retailers 2018. Accessed December ple audience context. J. Consumer Res. 32(2):260–265.
31, 2021, https://www.statista.com/forecasts/646030/united- Semuels A (2018) Is free shipping hurting Amazon? Accessed De-
states-top-online-stores-united-states-ecommercedb. cember 31, 2021, https://www.theatlantic.com/technology/
Newman JP, Ferguson ME, Garrow LA, Jacobs TL (2014) Estimation archive/2018/04/free-shipping-isnt-hurting-amazon/559052/.
of choice-based models using sales data from a single firm. Susarla A, Barua A, Whinston AB (2006) Understanding the service
Manufacturing Service Oper. Management 16(2):184–197. component of application service provision: An empirical anal-
Oliver R (1977) Effects of expectations and disconfirmation on post- ysis of satisfaction with ASP services. Information Systems Out-
exposure product evaluations. J. Appl. Psych. 62:246–250. sourcing (Springer, Berlin), 481–521.
Oliver RL (1980) A cognitive model of the antecedents and con- Terza JV, Basu A, Rathouz PJ (2008) Two-stage residual inclusion
sequences of satisfaction decisions. J. Marketing Res. 17(4): estimation: Addressing endogeneity in health econometric
460–469. modeling. J. Health Econom. 27(3):531–543.
Pagan A (1986) Two stage and related estimators and their applica- Train KE (2009) Discrete Choice Methods with Simulation (Cambridge
tions. Rev. Econom. Stud. 53(4):517–538. University Press, Cambridge, UK).
Park S, Gupta S (2012) Handling endogenous regressors by joint es- Tversky A, Kahneman D (1974) Judgment under uncertainty: Heu-
timation using copulas. Marketing Sci. 31(4):567–586. ristics and biases. Science 185(4157):1124–1131.
Perdikaki O, Kesavan S, Swaminathan JM (2012) Effect of traffic on Vulcano G, Van Ryzin G, Ratliff R (2012) Estimating primary de-
sales and conversion rates of retail stores. Manufacturing Service mand for substitutable products from sales transaction data.
Oper. Management 14(1):145–162. Oper. Res. 60(2):313–334.
Perez S (2018) Target launches free, 2-day shipping with no mini- Wang Z (2010) Anonymity, social image, and the competition for
mum purchase requirement. Accessed December 31, 2021, volunteers: A case study of the online market for reviews. B.E.
https://techcrunch.com/2018/10/24/target-launches-free-2- J. Econom. Anal. Policy 10(1).
day-shipping-with-no-minimum-purchase-requirement/. Wooldridge JM (2010) Econometric Analysis of Cross Section and Panel
Perez S (2019) Walmart announces next-day delivery on 200k+ Data.(MIT Press, Cambridge, MA).
items in select markets. Accessed December 31, 2021, https:// Wooldridge JM (2015) Control function methods in applied econo-
techcrunch.com/2019/05/13/walmart-announces-launch-of- metrics. J. Human Resources 50(2):420–445.
nextday-delivery-on-200k-items-in-select-markets. Wu F, Huberman BA (2008) How public opinion forms. Proc.
Richter F (2019) Infographic: Third-party sellers are outselling Internat. Workshop on Internet and Network Economics (Springer,
Amazon on Amazon. Accessed December 31, 2021, https:// Berlin), 334–341.